基于改进Grassberger熵随机森林分类器的目标检测

Grassberger entropy is improved, and the improved Grassberger entropy is used to compute information gain. The random forest classifier is trained by selecting the optimal split parameters of the split node. The trained random forest classifier predicts whether the proposal windows generated by selective search contain object. For each of training samples and proposal windows, one normalized gradient magnitude, three LUV color channels, and six histograms of oriented gradients are extracted. The algorithm performance is tested on SenseAndAvoid dataset, and the average detection precision of 73.2% is achieved. Results show that the average detection precision is more than 98% in the range of safety envelope. The improved Grassberger entropy computing information gain can promote precision of object detection.